Datasets for Detection and Localization of Speech Buried in Drone Noise

Abstract

This paper introduces datasets for detection and localization of speech buried in drone noise based on a scenario that a drone is used to search and rescue victims in disaster situations with microphones mounted in the drone. Since the distances between the blades and the microphones are short, the signal-to-noise ratio (SNR) can be exceptionally low. To emulate these signals, drone noise and impulse responses for convolving voice signals were recorded outdoor, and the voice signals were mixed in relatively low levels. The resultant SNRs of the present datasets was below -5 dB on average, which makes localization and detection fairly challenging. Detection and localization results with some baseline methods including a convolutional neural network (CNN) were presented. These datasets can be useful to develop and improve deep neural networks for rescue drones.